Be Columbus — Or you will never create an Autonomous Vehicle
(Alternate analogy — Disrupt the Distributors)
We all know the story (myth). Columbus believed the world was round. The rest of the world thought it was flat. That included the best minds of the time.
In current times the issue at hand is the use of public shadow and safety driving to develop and test autonomous vehicles. The world thinks this is the best or only way to do this. In this case the “world” comprises of Silicon Valley, IT and gaming industries as well as the most of the public, press, academics and government who believe they are the best engineering minds of our time on this subject matter.
Just as it seemed simple to prove Columbus was right, I believe it is just as easy to understand public shadow and safety driving can never result in anything close to a legitimate driverless vehicle and take lives needlessly trying. Just as the experts in Columbus’s day could come up with pretty simple use cases to resolve the issue, I believe we can do the same here by answering a couple questions objectively.
· Do you have the time and funding to finish? — How many “miles” do you have to drive and how much money would it cost to stumble and restumble on all the scenarios needed to train the extremely inefficient machine learning systems. RAND believes that is 500B at 10X better than a human. Toyota said it was equivalent to one trillion miles. Some very conservative math would show it would take over $300B to get through that trillion miles.
· How many people must be injured or killed to finish? — There are two significant safety issues here. The first involves when the system under test has to give back control of the vehicle to the human “safety driver when it can’t handle something. This is called handover or fallback. The problem here is that no monitor and alarm system can provide those safety drivers with the time to regain proper situational awareness in critical scenarios. The second issue is learning accident scenarios. In order for these systems to know how to avoid or handle accident scenarios they have to learn them like any other scenarios. This means experiencing thousands of them thousands of times each. This will cause thousands of injuries and deaths
· Aren’t there much more suitable alternatives? — This is where the problem really lies. Clearly test tracks would help in some areas but have similar issues as public roads. What this comes down to is simulation. The problem being that this “world” thinks simulation can’t replace the real-world. That of course is based on their exposure to gaming technology and the lack of understanding the first question I asked above. Of course, nothing can be more real than the real world. That however is not the issue. The issue is can you get what you need in the real world, repeat it thousands of times, and can simulation be real enough. The former is not remotely possible. That latter is and has been for decades. Only not in gaming but in the DOD and aerospace. With the help of real-world data to inform and validate the simulation the technology in DoD and aerospace, augmented by proper systems and safety engineering and the FAA simulation verification procedures as a guide, you can do enough of what is needed to get to that 10X better than a human level or more. The issue is not that a human has to be fooled in to thinking the simulation is real. What matters is if they, all the sensors, the perception, planning and execution systems can respond as if the simulation was real. The answer to that is it can. DoD has run simulated urban war games in the public environment now for a while. They not only have to deal with the same complexity as the public AV makers have to deal with, but their vehicles can drive off the road and folks are shooting at one another. Beyond that though is the most important part. DoD/aerospace/FAA simulation technology is far better than gaming technology. The gaming world made architecture decisions decades ago that severally limits their real-time, model fidelity and loading scaling capabilities. Those issues will lead to improperly trained systems, false confidence and real-world tragedies. (If would like me to demonstrate this for you please let me know.)
Given these issues are not that difficult to overcome, what is the real problem here? The real issue involves various and common machinations of ego, group dynamics and Maslov’s Triangle. This is a massive echo chamber. Made worse by the fact that most of the developers or engineers in these companies have little exposure to the type of engineering and simulation that is needed here, and they make far more money than in other domains. Many making $250k-$450k to start.
A simple look back at history with the aerospace industry being an example shows that the only way these dynamics are broken are increasingly specific laws and regulations. All of which are normally borne out of tragedies. The FAA was formed because of several tragedies in the 1950s. The driverless vehicle industry is going through the same thing. Four people have been killed needlessly by this approach. The death of Elaine Herzberg, the first non-driver and female, started the paradigm shift. The recent determination that Jeremy Banner died in the same manner as Joshua Brown 3 years prior and Walter Huang’s apparent refusal to settle and sue Tesla, stating these systems should be tested before they are allowed in the public domain. have greatly accelerated that shift. The question is what will it take to shift 99.9% of that public shadow driving to aerospace/DoD/FAA simulation technology? Will that have to include the first death of a child or family? In the past 3 months it is easy to see the shift accelerate. I have personally seen quite a few members of the press, public, industry participants and people from within the OEMs and autonomous vehicle makers begin to question things and suggest people should not be used as human Guinea pigs. If the industry is going to self-police and correct vs have the government mandate it for them, we will need for these progressive epiphanies to happen much faster and broader. We need for the majority to Be Columbus.
Please find more information here
Using the Real World is better than Proper Simulation for Autonomous Vehicle Development — NONSENSE
SAE Autonomous Vehicle Engineering Magazine-End Public Shadow Driving
Common Misconceptions about Aerospace/DoD/FAA Simulation for Autonomous Vehicles
The Hype of Geofencing for Autonomous Vehicles
